Yi of The University of Sydney

Abstract: Ecohydrology seeks to understand how water, energy, vegetation and carbon processes interact across space and time. With the rapid growth of satellite archives, environmental sensor networks and data-driven modelling, deep learning (DL) has become increasingly useful for estimating ecohydrological variables such as soil moisture, land surface temperature, evapotranspiration, vegetation water content, gross primary productivity and solar-induced chlorophyll fluorescence. These developments have made large-scale ecohydrological monitoring increasingly feasible, especially where direct observations are sparse. However, most DL applications remain strongly dependent on pattern recognition from the training data. This poses risks for ecohydrology, where target processes are highly dynamic, physically constrained and often non-stationary under climate variability, land-use change and extreme events. A purely data-driven model may perform well within the statistical envelope of its training data but become unreliable under unseen conditions. More importantly, ecohydrology requires not only prediction of what is happening, but also understanding of why it is happening through attribution, physical consistency and causal interpretation. Earth observation foundation models (EOFMs) offer a promising opportunity to refine spatiotemporal ecohydrological estimation by learning reusable representations from large, heterogeneous EO archives. Yet their role should not be viewed simply as larger-scale prediction. A key challenge is how EOFMs can be elegantly constrained by physical principles and integrated with process understanding. This talk will discuss recent progress in spatial, multimodal and temporal EOFMs, and examine how they may support physics-constrained estimation, uncertainty propagation, cross-sensor transfer and domain-specific benchmarking. Particular attention will be given to the need to distinguish lower-level EO-derived variables with stronger physical links from higher-level ecohydrological products built from multiple layers of proxies, assumptions and model constraints.

Biography:

Dr Yi Yu received a Ph.D. degree in hydrology and remote sensing from The Australian National University, Canberra, ACT, Australia. He is currently an Associate Lecturer with The University of Sydney, Sydney, NSW, Australia. His research interests focus on developing hybrid approaches that integrate physical and data-driven methods to better understand land–atmosphere interactions across various spatiotemporal scales, particularly in the context of climate extremes such as drought. Yi was a recipient of the Commonwealth Scientific and Industrial Research Organisation (CSIRO) Agriculture and Food Director’s Award—Next Generation Science in 2023. He was among the Student Paper Competition Finalists at the 2024 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). 

Email:

Address:The University of Sydney, , Sydney, Australia